NEURAL-NETWORK PREDICTION OF AE DATA

Neural network (NN) models were constructed to study prediction of the AE index. Both solar wind (vB(z)) and previous observed AE inputs wer e used to predict AE data for different numbers of time steps ahead. I t seems that prediction of the original unsmoothed AE data is possible only for 10 time steps (25 min) ahead. The predicted time series of t he AE data for 50 time steps (125 min) ahead was found to be dynamical ly different from the original time series. It is possible that the NN model cannot reproduce the turbulent part of the power spectrum of th e AE data. However, when using smoothed AE data the prediction for 10 time steps ahead gave an NMSE of 0.0438, and a correlation coefficient of 0.98. The predictive ability of the model gradually decreased as t he lead time of the predictions was increased, but was quite good up t o predictions for 30 time steps (75 min) ahead.